The Agentic Writing Frameworkthe writing process pros use — run as simple AI prompts
you a prompt a reviewer prompt automatic
the agentic writing framework · how it works

One prompt can't do eight jobs.
So every job got its own prompt.

Ask AI to write your ad in one prompt and you're asking it to find the hook, plan the structure, write, edit, and judge — all at once. No writer works that way. That's why one-shot AI copy reads like AI. This framework splits the job the way a writing team does: eleven steps, each one a simple prompt with one job, run in order, in the AI chat you already use.

Nothing here is software. Every box on the map is plain text you can read, copy, and change. Click any box to see what it does for your copy — and the exact prompt behind it.

How to run it

Three things to know before your first piece.

1 · Run the prompts in orderOpen a chat, paste the prompt, give it what it asks for. Carry each step's output into the next. Run the two review steps in a fresh chat — a reviewer who watched the draft being written can't read it cold.
2 · The brief does the heavy liftingFive minutes on the brief saves an hour of regenerating. Who's reading, what they already know, what you're selling, what you can prove, what you can never say. Vague in, vague out — every time.
3 · You're the only judgeNo step scores the copy out of ten — AI grading AI produces confident nonsense. The reviewers leave notes on exact lines. The last call is always yours.

How to make it yours

The prompts are the starting point. The system gets personal in three places.

Your brand voice fileOne page, filled in once per brand: who's speaking, how claims get made, five lines that sound exactly right, five that sound wrong, words you never use. Every writing prompt reads it — it's the difference between "AI copy" and copy that sounds like you.
Your edits become rulesEvery time you fix the finished piece, you're seeing your own taste in action. Notice the pattern — "I always cut the third adjective", "I never open with a question" — and add it to your voice file. The next draft arrives closer to done. Ten pieces in, it's writing in your voice.
The prompts are editableThey're plain text. Add a banned phrase. Change how many hooks you want. Tighten the beat structure for your format. If a step keeps making the same mistake for you, fix the prompt once instead of fixing the copy forever.

The line, start to finish

What each step does for your copy. Click any box on the map for the full story and the prompt itself.

StepWhat it does for your copyYou get
1 · Your briefGives the AI what a good freelancer would demand before writing a wordthe foundation every prompt works from
2 · ReceiptsLocks the facts: every claim you're allowed to make, with its proof — so nothing gets invented lateryour receipts sheet
3 · HooksSix genuinely different openings, tested against the five things every hook needs, ranked colda winner and a runner-up
4 · You pickYour call on the opening — or let the winner throughthe opening
5 · SkeletonPlans the piece like a pro: the right opening type for your reader, beats with jobs, proof placed where it landsthe beat-by-beat plan
6 · DraftWrites the piece in your voice, from your facts only — quotes stay word-for-wordthe first full draft
7 · The edit passOne disciplined pass for clarity, specificity, rhythm, flow — and the AI tells, swept outthe edited draft
GuardrailsAutomatic tripwires: no altered quotes, no banned phrases, not over length (optional)a pass, or a fix list
8 · The customer readYour buyer, cold: where they'd stop reading, what they don't believe, what's still unanswerednotes on specific lines
9 · The fact checkEvery sentence vs. your receipts — the reason you can publish without the 3am doubtnotes on specific lines
10 · Apply the notesFolds the notes in once, flags what it can't fix honestly, then stopsthe finished piece + the log
11 · Your callEdit, approve, or kill — and feed what you changed back into your voice filecopy you'd put your name on
The line only moves forward. No step re-runs itself and no AI grades AI in circles — when a step can't do something honestly, it flags it for you and keeps moving. One pass, about ten minutes, and then the only judge that matters: you.

The five things every hook needs

The first line decides everything, so it gets its own test. Shape is free — question, statement, story — the test is the elements, not the syntax. A hook missing one gets cut, not "improved".

Reading your reader always comes first. A formula that fights what your reader already knows loses to the reader, every time — and "make them feel the pain" is one line of recognition, never a paragraph.

Same facts, different hook

The difference between an opening that dies and one that works is almost never new information.

The rules that never bend

Everything else is taste. These aren't.

ruleNever invent a factIf it's not on your receipts sheet, it doesn't go in the copy. No invented testimonials, no rounded-up numbers, no features you don't have.
ruleNever upgrade the proof"8 people" never becomes "8 women" because it pulls harder. A trial stays a trial. The receipt is the ceiling.
ruleCustomer quotes are sacredReal quotes survive every step word-for-word. Only you may touch them.
ruleNo stacked fragments"Wide awake. Heart going." — fragments have no one in them. It's the fastest way to sound like AI trying to sound punchy.
ruleNo emotion by adjective"Amazing", "shocking", "game-changing" tell the reader what to feel — and make them feel nothing. Emotion comes from specifics.
ruleNo fake urgencyDeadlines that are real, scarcity that is real, or nothing. Readers can smell it, and it costs you every claim that comes after.

Why separate prompts beat one big prompt

The idea the whole framework stands on.

One job, done properlyA prompt asked to do everything averages everything. A prompt with one job — just hooks, just structure, just fact-checking — does that job at full strength. That's all "agentic" means here. Nothing more mysterious than that.
Fresh eyes are real eyesThe reviewers read the piece in a fresh chat, with no memory of why the draft says what it says. That's the closest AI gets to reading the way your customer reads: cold.
You can fix one partWeak hooks? Fix the hook prompt. Heavy-handed edits? Fix the edit prompt. One big prompt gives you one big dial. Eleven small ones give you a control room.

Words this system uses

Brief
The one-page setup: what you're selling, who's reading, how aware they are, what you want them to do, what you can prove, what's off-limits.
Raw material
Whatever the copy gets built from — your landing page, reviews, product sheet, research — pasted as-is, before any writing starts.
Receipts sheet
The numbered list of every claim you're allowed to make, each with its source quoted. The rule: no fact enters the copy that isn't on the list.
Keeper line
A real customer quote or a line that's already proven it works. Survives every step word-for-word.
Awareness
How much your reader already knows — from "doesn't know they have the problem" to "knows your product and wants a deal". It decides what your opening is allowed to say.
Twist
The moment your hook contradicts something the reader believes — or something they already tried that failed them. It's what opens the loop that keeps them reading.
Brand voice file
Your one-page description of how the brand talks: who's speaking, gold lines, banned words. Filled in once, read by every writing prompt, sharpened every time you edit.
Must-fix / should / minor
The only three labels a reviewer can put on a note. No scores — AI is honest about problems and dishonest about grades.

Get the framework

One zip. Plain-text prompts, your brand voice template, and a skill file that runs the whole line for you.

Set it up in Claude — recommended

  1. Download the zip. Don't unzip it — Claude handles that part.
  2. Open Claude Code in a folder. Go to claude.ai/code (or the Claude desktop app → Code) and open a folder. That's all a "project" is here — a folder Claude works inside. Use the one where you already keep your copy work, or make a brand-new empty one just for this. Both work; the framework brings everything else.
  3. Hand Claude the zip. Drag the zip file into the chat and say: install this skill. Claude creates the folders, unpacks it into .claude/skills/, and tells you when it's in.
  4. Run it. Type /agentic-writing-framework in the chat. It interviews you for the brief, runs all eleven steps in order — reviewers in fresh eyes — and stops for your call at the end.

Prefer to do it by hand? Unzip, then put the agentic-writing-framework folder inside .claude/skills/ in that same folder — create .claude/skills/ if it doesn't exist yet.

No Claude? Any AI chat works

Open the files in steps/ in order and paste each prompt with what it asks for. Run the two reviewers/ prompts in fresh chats — a reviewer that watched the draft being written can't read it cold.

First thing after installing

Fill in voices/_template.md — one page on how your brand talks. Every writing step reads it. It's the difference between AI copy and copy that sounds like you. (Installed via Claude? Just ask it to open the template for you.)

24 plain-text files · nothing to install · yours to edit

Watch the guide

Filip walks the whole system — what every step does and how to run your first piece. 8 minutes.